Add alternative relocation backends: GrowClust / GrowClust3D.jl / SPIDER

[seismoseo]

Motivation

PocketQuake currently offers two relocation backends, both double-difference least-squares (HypoDD-family): Fortran ph2dt + hypoDD (default), and Python relocDD-py (--python).

A different relative-relocation philosophy adds cross-validation and robustness — especially on sparse / noisy / weakly-linked clusters where HypoDD-family inversions produce "air-quakes" or fail to converge (e.g. the Namwon test, where a poorly-located outlier collapsed the dt.ct inversion).

Candidates — links + references

• GrowClust (Fortran90) — hierarchical clustering driven directly by cross-correlation, no damped matrix inversion; robust where least-squares DD struggles.

  • Code: https://github.com/dttrugman/GrowClust
  • Trugman, D. T., & Shearer, P. M. (2017). GrowClust: A Hierarchical Clustering Algorithm for Relative Earthquake Relocation, with Application to the Spanish Springs and Sheldon, Nevada, Earthquake Sequences. Seismological Research Letters, 88(2A), 379–391. (SCEC record: https://central.scec.org/publication/6649)

• GrowClust3D.jl (Julia) — the maintained successor; adds 3-D velocity-model support.

  • Code: https://github.com/dttrugman/GrowClust3D.jl
  • Trugman, D. T., Chamberlain, C. J., Savvaidis, A., & Lomax, A. (2023). GrowClust3D.jl: A Julia Package for the Relative Relocation of Earthquake Hypocenters Using 3D Velocity Models. Seismological Research Letters, 94(1), 443–456. https://doi.org/10.1785/0220220193

• SPIDER — Bayesian double-difference for whole catalogs via a physics-informed neural-network Eikonal solver + Stochastic Gradient Langevin Dynamics, parallelised over GPUs. Conceptually close to PocketQuake's HypoSVI locator (same Ross group), so it would pair naturally with the --python stack.

  • Ross, Z. E., Wilding, J. D., Azizzadenesheli, K., & Kato, A. (2026). SPIDER: Scalable Probabilistic Inference for Differential Earthquake Relocation. Journal of Geophysical Research: Solid Earth, 131, e2025JB032769. https://doi.org/10.1029/2025JB032769
  • Preprint: https://arxiv.org/abs/2508.12117 · Eos highlight: https://eos.org/editor-highlights/taming-the-seismicity-tsunami-with-a-scalable-bayesian-framework

Scope (rough)

• New --reloc-backend {growclust|...} selector (mirrors the hypodd / relocdd_py dispatch in hypodd.run_dtcc / pipeline.run_cluster).
• Feed it the existing dt.cc (xcorr stage) + station/event tables; emit a hypoDD.reloc-compatible product so 03_results_<cluster>.ipynb renders it.
• Decide how each method's uncertainties map onto the notebook's 95% error bars (GrowClust has its own bootstrap; SPIDER is natively Bayesian).

Out of scope

Not implementing now — tracking only.

— filed via Claude Code